Digital radiographic imaging detection prediction model training and process parameter optimization method

Abstract

The application relates to the technical field of digital imaging, in particular to a training method of a digital radiographic imaging detection prediction model and a process parameter optimization method, wherein the method comprises the following steps: constructing a prediction model of digital radiographic imaging detection; determining a nonlinear region of an implicit layer of the prediction model based on the output value variation range of the prediction model; setting a center point according to the nonlinear region; and mixing a Gaussian kernel and an inverse radial kernel function to fit the nonlinear region; generating a training data set according to an orthogonal test data set of digital radiographic imaging to train the prediction model; in the training process, optimizing nodes of the implicit layer according to the mean square error and the prediction accuracy of samples in the training data set; and testing the prediction model by using random combination parameters to construct a test set until the prediction performance of the prediction model reaches a target requirement. Therefore, the problems that the reliability of a model for optimizing process parameters of digital radiographic imaging detection is poor and the fitting capacity is limited in the related art are solved.

Description

Technical Field

[0001] This application relates to the field of digital imaging technology, and in particular to a method for training and optimizing process parameters of a digital X-ray imaging detection prediction model. Background Technology

[0002] In the field of industrial nondestructive testing, digital radiographic imaging is widely used for quality assessment of welds, castings, composite materials, and other materials. To obtain high-quality radiographic images, key process parameters need to be precisely adjusted. The optimal process parameters are not fixed for different materials, thicknesses, and testing requirements; rather, they need to be optimized based on specific testing conditions. Currently, the core objective of process parameter optimization is to improve testing efficiency and reduce unnecessary exposure time and radiation dose while ensuring image quality.

[0003] Related technologies employ multi-objective response surface regression models to optimize process parameters. These models use second-order polynomial regression to calculate the impact of parameters on image quality and solve for the optimal combination. However, the simulation's fitting ability is limited. Multi-objective response surface regression models, primarily based on second-order polynomial models, struggle to effectively handle high-dimensional and complex data, especially under conditions of strong nonlinearity. This limits optimization accuracy and places high demands on the distribution of experimental data to prevent affecting model fitting. However, in actual testing, experimental data is often uneven, leading to reduced reliability of the optimized model. Summary of the Invention

[0004] This application provides a method for training a digital X-ray imaging detection prediction model and optimizing process parameters, in order to solve the problems of poor reliability and limited fitting ability of models used to optimize digital X-ray imaging detection process parameters in related technologies.

[0005] The first aspect of this application provides a training method for a digital ray imaging detection prediction model, comprising the following steps: constructing a prediction model for digital ray imaging detection, wherein the prediction model includes an input layer, an output layer, a hidden layer, and an activation function; the hidden layer determines a nonlinear region based on the magnitude of the output value change of the prediction model, sets a center point according to the nonlinear region, and the activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region; obtaining an orthogonal experimental dataset for digital ray imaging, generating a training dataset based on the orthogonal experimental dataset, training the prediction model using the training dataset, and optimizing the nodes of the hidden layer according to the mean square error and prediction accuracy of the samples in the training dataset during the training process; constructing a test set using randomly combined parameters, testing the prediction model using the test set, and updating the network parameters of the prediction model according to the test results of the prediction model until the prediction performance of the prediction model reaches the target requirements.

[0006] Optionally, before training the prediction model using the training dataset, the process includes: selecting initial center points based on a sample density weighting strategy; identifying target samples whose output value changes by a greater than a preset magnitude, and increasing the weight of the target samples to increase the number of center points in the nonlinear region of the hidden layer.

[0007] Optionally, the weights of the target samples are:

[0008]

[0009] Among them, w i Let Δy be the weight of the i-th target sample. i Let Δy be the magnitude of the change in the output value of the i-th target sample. j Let N be the range of change in the output value of the j-th sample, N be the total number of samples, and j be the sample index variable.

[0010] The loss function of the prediction model is:

[0011]

[0012] Where N is the total number of samples, w i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. i Let L be the true value of the i-th sample, and L be the total loss of the prediction model.

[0013] Optionally, after setting the center point according to the nonlinear region, the method further includes: identifying the distribution density of samples near each center point; and adjusting the width parameter of the corresponding center point based on the distribution density.

[0014] Optionally, the prediction model is trained using the training dataset, including: training the prediction model using samples in the training dataset and obtaining the mean squared error of the prediction model; calculating the local gradient of the prediction model based on the mean squared error; and adjusting the center point and center point width parameters of the hidden layer using the local gradient.

[0015] Optionally, obtaining an orthogonal experimental dataset for digital X-ray imaging includes: obtaining multiple input parameters for digital X-ray imaging, wherein the input parameters are detection process parameters; determining a target orthogonal array based on the multiple input parameters and multiple levels; combining the multiple input parameters based on the target orthogonal array to obtain multiple sets of input parameters; performing digital X-ray imaging on the multiple sets of input parameters to obtain corresponding output parameters, wherein the output parameters are image quality evaluation indicators; and constructing an orthogonal experimental dataset based on the multiple sets of input parameters and the corresponding output parameters.

[0016] A second aspect of this application provides a method for optimizing digital X-ray imaging detection process parameters. The method optimizes the prediction model trained using the training method of the digital X-ray imaging detection prediction model described in the above embodiments, and includes the following steps: obtaining the target image quality index input by the user; and optimizing the digital X-ray imaging detection process parameters that meet the target image quality index based on the target image quality index and the prediction model.

[0017] A third aspect of this application provides a training apparatus for a digital ray imaging detection prediction model, comprising: a construction module for constructing a prediction model for digital ray imaging detection, wherein the prediction model includes an input layer, an output layer, a hidden layer, and an activation function; the hidden layer determines a nonlinear region based on the magnitude of the output value change of the prediction model, sets a center point according to the nonlinear region, and the activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region; a training module for acquiring an orthogonal experimental dataset for digital ray imaging, generating a training dataset based on the orthogonal experimental dataset, training the prediction model using the training dataset, and optimizing the nodes of the hidden layer according to the mean square error and prediction accuracy of the samples in the training dataset during the training process; and an update module for constructing a test set using randomly combined parameters, testing the prediction model using the test set, and updating the network parameters of the prediction model according to the test results of the prediction model until the prediction performance of the prediction model reaches the target requirement.

[0018] Optionally, it also includes: adding a module for selecting initial center points based on a sample density weighting strategy before training the prediction model using the training dataset; identifying target samples whose output value changes by a greater than a preset magnitude, and increasing the weight of the target samples to increase the number of center points in the nonlinear region of the hidden layer.

[0019] Optionally, the weights of the target samples are:

[0020]

[0021] Among them, w i Let Δy be the weight of the i-th target sample. i Let Δy be the magnitude of the change in the output value of the i-th target sample. j Let N be the range of change in the output value of the j-th sample, N be the total number of samples, and j be the sample index variable.

[0022] The loss function of the prediction model is:

[0023]

[0024] Where N is the total number of samples, w i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. iLet L be the true value of the i-th sample, and L be the total loss of the prediction model.

[0025] Optionally, it also includes: an adjustment module for identifying the distribution density of samples near each center point after setting the center point according to the nonlinear region; and adjusting the width parameter of the corresponding center point based on the distribution density.

[0026] Optionally, the training module is further used to: train a prediction model using samples in the training dataset and obtain the mean squared error of the prediction model; calculate the local gradient of the prediction model based on the mean squared error; and adjust the center point and center point width parameters of the hidden layer using the local gradient.

[0027] Optionally, the training module is further used to: obtain multiple input parameters for digital X-ray imaging, wherein the input parameters are detection process parameters; determine a target orthogonal array based on the multiple input parameters and multiple levels, and combine the multiple input parameters based on the target orthogonal array to obtain multiple sets of input parameters; perform digital X-ray imaging on the multiple sets of input parameters to obtain corresponding output parameters, wherein the output parameters are image quality evaluation indicators; and construct an orthogonal experimental dataset based on the multiple sets of input parameters and the corresponding output parameters.

[0028] The fourth aspect of this application provides a digital X-ray imaging detection process parameter optimization device, which optimizes the prediction model trained by the training device of the digital X-ray imaging detection prediction model in the above embodiments, including: an acquisition module for acquiring a target image quality index input by a user; and an optimization module for optimizing digital X-ray imaging detection process parameters that conform to the target image quality index based on the target image quality index and the prediction model.

[0029] A fifth aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to perform a training method for a digital X-ray imaging detection prediction model as described above, or a method for optimizing digital X-ray imaging detection process parameters.

[0030] Therefore, this application has at least the following beneficial effects:

[0031] This application embodiment can construct a prediction model for digital X-ray imaging detection. The hidden layer of the prediction model determines the nonlinear region based on the change range of the prediction model's output value. A center point is set according to the nonlinear region. The activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region, thereby improving the model's fitting ability in asymmetric nonlinear regions. The prediction model is trained using a training dataset generated by orthogonal experimental datasets. The training dataset generated by orthogonal experimental datasets can ensure that the samples are uniformly distributed and the variables are not collinear, avoiding the limitations of traditional response surface methodology which has strict requirements on data distribution. During the training process, the nodes of the hidden layer are continuously optimized to improve the prediction accuracy and reliability of the model. Then, a test set is constructed using random combination parameters to test the prediction model. The network parameters of the prediction model are updated according to the test results until the prediction performance of the prediction model reaches the target requirements. The test set constructed by random combination parameters can ensure that its sample distribution is different from that of the training dataset, so as to more accurately evaluate the model's generalization ability and prediction performance. The constructed training prediction model can be more adaptable to the optimization of nonlinear digital X-ray imaging detection process parameters in the future, thereby improving the reliability of subsequent optimization results. This solves the technical problems in related technologies, such as the poor reliability and limited fitting ability of models used to optimize digital X-ray imaging detection process parameters.

[0032] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0033] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0034] Figure 1 This is a flowchart of a training method for a digital ray imaging detection prediction model provided according to an embodiment of this application;

[0035] Figure 2 This is a structural diagram of the prediction model provided according to the embodiments of this application;

[0036] Figure 3 A flowchart of a method for optimizing digital X-ray imaging detection process parameters according to an embodiment of this application;

[0037] Figure 4 This is a schematic diagram of a training device for a digital ray imaging detection prediction model provided according to an embodiment of this application;

[0038] Figure 5 A schematic diagram of a digital X-ray imaging detection process parameter optimization device provided in the embodiments of this application;

[0039] Figure 6This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0040] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0041] Before describing the solution of this application, the following related technologies are introduced. The related technologies for digital X-ray imaging detection process parameters mainly include the following methods.

[0042] 1. Determine the initial parameters by finding the exposure curve. The exposure curve is usually obtained from empirical data or experiments, and provides recommended ranges of process parameters for different materials, thicknesses and testing processes.

[0043] However, the following problems exist: it is highly dependent on expert experience, and professional personnel need to conduct repeated experiments to find relatively suitable process parameters. Multiple exposures are required during the experiment, which increases the detection time and the radiation dose, affecting detection efficiency and safety. Operators make fine adjustments based on the parameters provided by the exposure curve until acceptable image quality is obtained. For complex working conditions, such as multi-layered materials and workpieces of different shapes, the trial and error process is more complicated and it is difficult to guarantee the optimal result.

[0044] Furthermore, exposure curves only provide an empirical range of parameters, not necessarily optimal ones. Different equipment, detection environments, and material properties all affect image quality, meaning exposure curves cannot fully adapt to all detection needs. Since global optimization is impossible, better parameter combinations may exist, but traditional methods struggle to accurately obtain them. Exposure curves are typically built based on simple linear or low-order nonlinear regression methods, which are insufficient for complex high-order nonlinear relationships. When multiple parameters (voltage, tube current, focal length, etc.) simultaneously affect image quality, exposure curves cannot accurately represent the complex relationships between these variables.

[0045] 2. The process parameters are optimized using a multi-objective response surface regression model. The impact of the parameters on image quality is calculated using a second-order polynomial regression model, and the optimal combination is solved.

[0046] However, the following problems exist: limited fitting ability. The multi-objective response surface regression model is mainly based on the second-order polynomial model, which is difficult to effectively handle high-dimensional complex data, especially when the nonlinear characteristics are strong, thus limiting the optimization accuracy; high requirements for experimental data distribution. Statistical methods require that the experimental data be evenly distributed, otherwise the model fitting effect will drop significantly. In actual testing, the experimental data is often uneven, which leads to a decrease in the reliability of the optimization model.

[0047] To address this, this application provides a training method for a digital X-ray imaging detection prediction model. In this method, a prediction model for digital X-ray imaging detection is constructed. The hidden layer of the prediction model determines the nonlinear region based on the magnitude of the output value change. A center point is set according to the nonlinear region. A mixture of Gaussian and inverse radial kernel functions is used to fit the nonlinear region, improving the model's fitting ability in asymmetric nonlinear regions. The prediction model is trained using a training dataset generated from an orthogonal experimental dataset. This orthogonal dataset ensures uniform sample distribution and non-collinear variables, avoiding the limitations of traditional response surface methodology's stringent requirements on data distribution. During training, the nodes of the hidden layer are continuously optimized to improve the model's prediction accuracy and reliability. A test set is then constructed using randomly combined parameters to test the prediction model. The network parameters of the prediction model are updated based on the test results until the prediction performance meets the target requirements. The test set constructed using randomly combined parameters ensures that its sample distribution differs from the training dataset, allowing for more accurate evaluation of the model's generalization ability and prediction performance. The resulting trained prediction model is better adapted to the optimization of nonlinear digital X-ray imaging detection process parameters, improving the reliability of subsequent optimization results.

[0048] Specifically, Figure 1 This is a flowchart illustrating a training method for a digital ray imaging detection prediction model provided in an embodiment of this application.

[0049] like Figure 1 As shown, the training method for this digital X-ray imaging detection prediction model includes the following steps:

[0050] In step S101, a prediction model for digital X-ray imaging detection is constructed. The prediction model includes an input layer, an output layer, a hidden layer, and an activation function. The hidden layer determines the nonlinear region based on the change amplitude of the output value of the prediction model and sets the center point according to the nonlinear region. The activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region.

[0051] The nonlinear region is the area where the input and output parameters exhibit a complex nonlinear relationship; the center point is the center vector of the kernel function in the hidden layer, which determines the model's ability to partition the input space.

[0052] The prediction model in this application embodiment can be modeled based on an RBF (Radial Basis Function) neural network, and some improvements have been made to the structure to better adapt it to the high-dimensional nonlinear process optimization problem of multiple process parameters in digital X-ray imaging inspection. The structure of the prediction model is as follows: Figure 2 As shown, the model includes an input layer, hidden layer 1 (i.e., hidden layer 1), hidden layer 2, and an output layer. The input layer takes into account the process parameters of digital X-ray imaging detection, including voltage, current, time, focal length, and frame rate. The output layer outputs image quality evaluation indicators, including sensitivity, spatial resolution, signal-to-noise ratio, and image grayscale value. The hidden layer of the prediction model can determine the nonlinear region based on the magnitude of the output value change and set the center point according to the nonlinear region to improve the local fitting ability of the model. Furthermore, the activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region, thereby improving the model's fitting ability in asymmetric nonlinear regions.

[0053] In step S102, an orthogonal experimental dataset of digital ray imaging is obtained, a training dataset is generated based on the orthogonal experimental dataset, and the prediction model is trained using the training dataset. During the training process, the nodes of the hidden layer are optimized based on the mean square error and prediction accuracy of the samples in the training dataset.

[0054] It is understood that the embodiments of this application can generate a training dataset based on the orthogonal experimental dataset of digital ray imaging, and use the training dataset to train the prediction model. During the training process, the nodes of the hidden layer are optimized according to the mean square error and prediction accuracy of the samples in the training dataset. The training dataset generated by the orthogonal experimental dataset can ensure that the samples are uniformly distributed and the variables are not collinear, avoiding the limitations of the traditional response surface method which has strict requirements on data distribution. Furthermore, the nodes of the hidden layer are continuously optimized during the training process to improve the prediction accuracy of the model.

[0055] Specifically, the number of nodes in the hidden layer does not use fixed empirical values, but is based on the "structural increment method". Starting with a small number of nodes (such as 5), the number is gradually increased, and the changes in mean square error and prediction accuracy are monitored. The number of nodes with the smallest error and the strongest generalization ability is selected, which is usually in the range of 10 to 25.

[0056] In this embodiment of the application, a training dataset is generated based on the orthogonal experimental dataset. The data in the orthogonal experimental dataset can be normalized (in the range of 0 to 1) to obtain the training dataset.

[0057] In this embodiment of the application, obtaining an orthogonal experimental dataset for digital X-ray imaging includes: obtaining multiple input parameters for digital X-ray imaging, wherein the input parameters are detection process parameters; determining a target orthogonal array based on the multiple input parameters and multiple levels; combining the multiple input parameters based on the target orthogonal array to obtain multiple sets of input parameters; performing digital X-ray imaging on the multiple sets of input parameters to obtain corresponding output parameters, wherein the output parameters are image quality evaluation indicators; and constructing an orthogonal experimental dataset based on the multiple sets of input parameters and the corresponding output parameters.

[0058] The input parameters are detection process parameters, including voltage, current, time, focal length, and frame rate; the output parameters are image quality evaluation indicators, including sensitivity, spatial resolution, signal-to-noise ratio, and image grayscale value; the target orthogonal array can be L... 36 Orthogonal array or L 25 Orthogonal array.

[0059] It is understood that the embodiments of this application can obtain multiple input parameters of digital ray imaging, determine a target orthogonal array based on multiple input parameters and multiple levels, combine multiple input parameters based on the target orthogonal array to obtain multiple sets of input parameters, perform digital ray imaging on each set of input parameters to obtain the corresponding output parameters, and then construct an orthogonal experimental dataset based on multiple sets of input parameters and corresponding output parameters.

[0060] Specifically, the embodiments of this application can employ orthogonal experimental design to construct a representative and comprehensive training sample set, ensuring uniform sample distribution and non-collinear variables. This avoids the limitations of traditional response surface methodology, which has stringent requirements on data distribution, and also avoids prediction instability caused by sparsity of experimental data. Taking a 5-factor, 6-level model as an example, L... 36 (66) Orthogonal arrays are used to ensure that the multidimensional variable space is covered with a small number of experiments. Each set of parameters corresponds to one actual imaging experiment. The above four image quality indicators are collected to form input-output pairs.

[0061] In this embodiment of the application, training the prediction model using the training dataset includes: training the prediction model using samples in the training dataset and obtaining the mean squared error of the prediction model; calculating the local gradient of the prediction model based on the mean squared error; and adjusting the center point and center point width parameters of the hidden layer using the local gradient.

[0062] It is understood that the embodiments of this application can use samples in the training dataset to train the prediction model, and calculate the local gradient of the prediction model based on the mean square error of the prediction model. The local gradient is used to adjust the center point and center point width parameters of the hidden layer to maintain the advantage of linear solvability of the model output layer weights, while also enhancing the adaptability of the model.

[0063] Specifically, this application introduces a local feedback adjustment mechanism. After the initial training is completed, the center point and width parameters are fine-tuned by the output error (similar to the local adjustment in the backpropagation algorithm process), which maintains the advantage of linear solvability of the output layer weights and enhances the model's predictive ability in the highly nonlinear response region.

[0064] In this embodiment of the application, before training the prediction model using the training dataset, the method includes: selecting initial center points based on a sample density weighting strategy; identifying target samples whose output value changes by a greater than a preset magnitude, and increasing the weight of the target samples to increase the number of center points in the nonlinear region of the hidden layer.

[0065] The preset range can be set according to specific circumstances, and there are no specific limitations on it.

[0066] It is understood that the embodiments of this application can select the initial center point based on the sample density weighting strategy and identify samples with large output value changes to increase their weight, so that the prediction model can have more center points in the highly nonlinear region.

[0067] In this embodiment of the application, the weight of the target sample is:

[0068]

[0069] Among them, w i Let Δy be the weight of the i-th target sample. i Let Δy be the magnitude of the change in the output value of the i-th target sample. j Let N be the range of change in the output value of the j-th sample, N be the total number of samples, and j be the sample index variable.

[0070] The loss function of the prediction model is:

[0071]

[0072] Where N is the total number of samples, w i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. i Let L be the true value of the i-th sample, and L be the total loss of the prediction model.

[0073] In this embodiment of the application, after setting the center point according to the nonlinear region, the method further includes: identifying the distribution density of samples near each center point; and adjusting the width parameter of the corresponding center point based on the distribution density.

[0074] It is understood that the embodiments of this application can also identify the distribution density of samples near each center point, and adjust the width parameter of the corresponding center point based on the distribution density to avoid overfitting or underfitting of the model caused by a globally uniform width parameter.

[0075] In step S103, a test set is constructed using randomly combined parameters. The prediction model is then tested using the test set. The network parameters of the prediction model are updated based on the test results until the prediction performance of the prediction model meets the target requirements.

[0076] The target requirements can be set according to specific circumstances and are not specifically limited. For example, if the prediction accuracy reaches a certain performance level, the main evaluation indicators may include MSE (mean squared error), MAPE (mean absolute percentage error), and R. 2 (Determination coefficient); Network parameters include the number of hidden layer nodes, the center point of the hidden layer, and the width of the center point of the hidden layer.

[0077] It is understood that the embodiments of this application can use random combination of parameters to construct a test set, and use the test set to test the prediction model. The network parameters of the prediction model are updated according to the test results of the prediction model until the prediction performance of the prediction model reaches the target requirements. The test set constructed by random combination of parameters can ensure that its sample distribution is different from that of the training dataset, so as to more accurately evaluate the generalization ability and prediction performance of the model.

[0078] The following specific embodiment describes the construction and training process of the digital ray imaging detection prediction model of this application, including:

[0079] Step 1: Define the input and output parameters, as shown in Table 1.

[0080] Table 1

[0081] Input parameters (process parameters) Output parameters (image quality evaluation metrics) Transmission voltage (kV) Sensitivity Tube current (mA) Spatial resolution Focal length (mm) Signal-to-noise ratio Single frame exposure time (s) Image grayscale value Overlay Frames

[0082] Step 2: Design orthogonal experiments and construct training datasets.

[0083] 1. An orthogonal experimental design method is adopted to construct a representative and comprehensive training sample set, ensuring that the samples are uniformly distributed and the variables are not collinear, thus avoiding the limitations of the traditional response surface methodology which has strict requirements on data distribution.

[0084] 2. Taking 5 factors and 6 levels as an example, L is selected. 36 (66) Orthogonal arrays ensure that a multidimensional variable space is covered with a small number of experiments.

[0085] 3. Each set of parameters corresponds to one actual imaging experiment, and the above four image quality indicators are collected to form an input-output pair.

[0086] 4. All experimental data were normalized (range 0 to 1) to facilitate subsequent network training.

[0087] Step 3: Construct an improved RBF neural network model (i.e., a prediction model).

[0088] To improve the fitting ability and stability of the RBF neural network in the DR process parameter prediction task, this application makes the following improvements to the traditional RBF model:

[0089] 1. Customized network structure.

[0090] Number of input layer nodes: consistent with the number of process parameters (i.e., 5); Number of output layer nodes: consistent with the image quality evaluation index (i.e., 4).

[0091] Hidden layer node number optimization: Instead of using a fixed empirical value, it is based on the "structural increment method", starting from a small number of nodes (such as 5) and gradually increasing it, monitoring the changes in mean square error (MSE) and prediction accuracy, and selecting the number of nodes with the smallest error and the strongest generalization ability, which is usually in the range of 10 to 25.

[0092] 2. RBF center point optimization.

[0093] In traditional RBF (Recurrent Baseline Function) clustering, hidden layer centers are typically selected randomly or by K-means clustering. This application employs an improved K-means algorithm, with improvements including:

[0094] The initial centroid selection is based on a sample density weighting strategy to avoid the sensitivity of K-means to initial values.

[0095] For samples with drastic changes in output variables, higher weights are given so that RBF has more center points in highly nonlinear regions.

[0096] Specifically, calculate the output gradient or the magnitude of change in the output value for each sample, Δy. i =y i -y i-1 Normalize the gradient to the range [0,1]:

[0097] When training the RBF network, this weight is incorporated into the loss function.

[0098]

[0099] This is equivalent to making samples with large gradients "speak louder" during network training, thereby guiding the network to place more center points in these areas and improving its nonlinear fitting ability.

[0100] 3. Activation function optimization.

[0101] Traditional RBF uses a Gaussian kernel function. This application dynamically adjusts σ (width parameter): σ is set individually for each center point.i Automatically adjusted based on the distribution density of its neighboring samples: σ i ∝1 / local_density(c i To avoid overfitting or underfitting of the model due to a globally uniform σ,

[0102] Introducing a multi-kernel activation mechanism: Using a combination of Gaussian kernels and inverse radial kernel functions (e.g., inverse multiquadric) to improve the model's fitting ability in asymmetric nonlinear regions.

[0103] 4. Training algorithm optimization.

[0104] In traditional RBF training, a fixed two-stage training method is used: first clustering centers, then linear weighting. This application introduces a local gradient feedback adjustment mechanism. After the initial training is completed, the center point and width parameters are fine-tuned through the output error (similar to the local adjustment in the BP process), maintaining the advantage of linear solvability of the output layer weights, while enhancing the model's adaptability.

[0105] Step 4: Model training and validation.

[0106] 1. Train the RBF model using an orthogonal experimental dataset. The test set is constructed using a random combination of parameters to ensure that its sample distribution is different from that of the training set, and is used to evaluate the model's generalization ability.

[0107] 2. Using MSE (mean squared error), MAPE (mean absolute percentage error), and R... 2 The coefficient of determination is the main evaluation indicator to ensure that the model accuracy is better than that of the traditional second-order regression model.

[0108]

[0109] 3. If the prediction accuracy is not up to standard, iterative optimization should be performed according to the "structural increment method" or the fine-tuning center point strategy.

[0110] According to the training method of the digital X-ray imaging detection prediction model proposed in the embodiments of this application, a prediction model for digital X-ray imaging detection can be constructed. The hidden layer of the prediction model determines the nonlinear region based on the change amplitude of the output value of the prediction model. The center point is set according to the nonlinear region. The activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region, so as to improve the model's fitting ability in asymmetric nonlinear regions. The prediction model is trained through a training dataset generated by orthogonal experimental dataset. The training dataset generated by orthogonal experimental dataset can ensure that the samples are uniformly distributed and the variables are not collinear, avoiding the limitation of the traditional response surface method which has strict requirements on data distribution. During the training process, the nodes of the hidden layer are continuously optimized to improve the prediction accuracy and reliability of the model. Then, a test set is constructed by randomly combining parameters to test the prediction model. The network parameters of the prediction model are updated according to the test results of the prediction model until the prediction performance of the prediction model reaches the target requirements. The test set constructed by randomly combining parameters can ensure that its sample distribution is different from that of the training dataset, so as to more accurately evaluate the generalization ability and prediction performance of the model. The constructed training prediction model can be more adaptable to the optimization of nonlinear digital X-ray imaging detection process parameters in the future, improving the reliability of subsequent optimization results.

[0111] This application also provides a method for optimizing process parameters of digital X-ray imaging detection, which optimizes the prediction model trained by the training method of the digital X-ray imaging detection prediction model in the above embodiments.

[0112] like Figure 3 As shown, the method for optimizing digital X-ray imaging detection process parameters includes the following steps:

[0113] In step S201, the target image quality index input by the user is obtained.

[0114] Image quality metrics can include the aforementioned sensitivity, spatial resolution, signal-to-noise ratio, and image grayscale value, specifically, sensitivity of 2%, signal-to-noise ratio >25dB, and grayscale value within the range of 180±10.

[0115] In step S202, the digital X-ray imaging detection process parameters that meet the target image quality index are optimized based on the target image quality index and the prediction model.

[0116] It is understood that the embodiments of this application can optimize digital X-ray imaging detection process parameters that meet the target image quality index based on the target image quality index and the prediction model. The prediction model infers the most matching detection process parameters based on the target image quality index, without relying on expert experience, thus improving optimization efficiency. Moreover, the target image quality index can include multiple indicators, enabling multi-objective optimization to improve image quality, specifically:

[0117] The prediction model achieves inverse solution through the inverse mapping algorithm, that is, given the target index y*, the solution is to find the parameter combination x such that f(x) = y*, where f is the trained prediction model, and the best matching detection process parameters are selected from multiple solutions.

[0118] Specifically, the digital X-ray imaging detection process parameter optimization method of this application includes:

[0119] 1. User-input image quality targets (e.g., sensitivity <2%, signal-to-noise ratio >25dB, grayscale value in the range of 180±10);

[0120] 2. The prediction model automatically back-calculates the most suitable combination of process parameters (i.e., illumination voltage, tube current, etc.);

[0121] 3. If multiple combinations meet the requirements, the combination with the shortest exposure time or the highest signal-to-noise ratio can be selected by further combining the principle of "maximum detection efficiency" (secondary optimization).

[0122] According to the digital X-ray imaging detection process parameter optimization method proposed in the embodiments of this application, the target image quality index input by the user can be obtained, and the prediction model can back-infer the most matching detection process parameters based on the target image quality index. This method does not rely on expert experience, improves optimization efficiency, and the target image quality index can include multiple indicators, enabling multi-objective optimization to improve image quality.

[0123] Next, referring to the accompanying drawings, a training device for a digital ray imaging detection prediction model and a device for optimizing digital ray imaging detection process parameters are described according to embodiments of this application.

[0124] Figure 4 This is a block diagram of a training device for a digital ray imaging detection prediction model according to an embodiment of this application.

[0125] like Figure 4 As shown, the training device 10 for the digital ray imaging detection prediction model includes: a construction module 101, a training module 102, and an update module 103.

[0126] The construction module 101 is used to construct a prediction model for digital X-ray imaging detection. The prediction model includes an input layer, an output layer, a hidden layer, and an activation function. The hidden layer determines the nonlinear region based on the change amplitude of the prediction model's output value and sets the center point according to the nonlinear region. The activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region. The training module 102 is used to obtain an orthogonal experimental dataset for digital X-ray imaging, generate a training dataset based on the orthogonal experimental dataset, and train the prediction model using the training dataset. During the training process, the nodes of the hidden layer are optimized based on the mean square error and prediction accuracy of the samples in the training dataset. The update module 103 is used to construct a test set using randomly combined parameters, test the prediction model using the test set, and update the network parameters of the prediction model based on the test results until the prediction performance of the prediction model meets the target requirements.

[0127] In this embodiment of the application, the training device 10 for the digital ray imaging detection prediction model further includes an addition module.

[0128] The module adds a component to select initial center points based on a sample density weighting strategy before training the prediction model using the training dataset; it also identifies target samples whose output value changes by a greater than a preset amount and increases the weight of these target samples to increase the number of center points in the nonlinear region of the hidden layer.

[0129] In this embodiment of the application, the weight of the target sample is:

[0130]

[0131] Among them, w i Let Δy be the weight of the i-th target sample. i Let Δy be the magnitude of the change in the output value of the i-th target sample. j Let N be the range of change in the output value of the j-th sample, N be the total number of samples, and j be the sample index variable.

[0132] The loss function of the prediction model is:

[0133]

[0134] Where N is the total number of samples, w i Let i be the weight of the i-th sample. Let y be the predicted value of the i-th sample. i Let L be the true value of the i-th sample, and L be the total loss of the prediction model.

[0135] In this embodiment of the application, the training device 10 for the digital ray imaging detection prediction model further includes an adjustment module.

[0136] The adjustment module is used to identify the distribution density of samples near each center point after setting the center point according to the nonlinear region; and adjust the width parameter of the corresponding center point based on the distribution density.

[0137] In this embodiment, the training module 102 is further configured to: train a prediction model using samples in the training dataset and obtain the mean squared error of the prediction model; calculate the local gradient of the prediction model based on the mean squared error; and adjust the center point and center point width parameters of the hidden layer using the local gradient.

[0138] In this embodiment, the training module 102 is further configured to: acquire multiple input parameters for digital X-ray imaging, wherein the input parameters are detection process parameters; determine a target orthogonal array based on the multiple input parameters and multiple levels; combine the multiple input parameters based on the target orthogonal array to obtain multiple sets of input parameters; perform digital X-ray imaging on the multiple sets of input parameters to obtain corresponding output parameters, wherein the output parameters are image quality evaluation indicators; and construct an orthogonal experimental dataset based on the multiple sets of input parameters and the corresponding output parameters.

[0139] It should be noted that the explanation of the training method embodiment for the digital ray imaging detection prediction model described above also applies to the training device for the digital ray imaging detection prediction model in this embodiment, and will not be repeated here.

[0140] The training device for the digital X-ray imaging detection prediction model proposed in this application can construct a prediction model for digital X-ray imaging detection. The hidden layer of the prediction model determines the nonlinear region based on the magnitude of the output value change. A center point is set according to the nonlinear region. The activation function uses a mixture of Gaussian kernels and inverse radial kernels to fit the nonlinear region, thereby improving the model's fitting ability in asymmetric nonlinear regions. The prediction model is trained using a training dataset generated from an orthogonal experimental dataset. This training dataset ensures uniform sample distribution and non-collinear variables, avoiding the limitations of traditional response surface methodology's stringent requirements on data distribution. During training, the nodes of the hidden layer are continuously optimized to improve the model's prediction accuracy. A test set is then constructed using randomly combined parameters to test the prediction model. The network parameters of the prediction model are updated based on the test results until the prediction performance meets the target requirements. The test set constructed using randomly combined parameters ensures that its sample distribution differs from the training dataset, allowing for more accurate evaluation of the model's generalization ability and prediction performance. Therefore, the constructed training prediction model can better adapt to the optimization of nonlinear digital X-ray imaging detection process parameters, improving the reliability of subsequent optimization results.

[0141] Figure 5 This is a block diagram of a digital X-ray imaging detection process parameter optimization device according to an embodiment of this application.

[0142] like Figure 5 As shown, the digital X-ray imaging detection process parameter optimization device 20 optimizes the prediction model trained by the aforementioned digital X-ray imaging detection prediction model training device, and includes: an acquisition module 201 and an optimization module 202.

[0143] The acquisition module 201 is used to acquire the target image quality index input by the user; the optimization module 202 is used to optimize the digital X-ray imaging detection process parameters that meet the target image quality index based on the target image quality index and the prediction model.

[0144] The digital X-ray imaging detection process parameter optimization device proposed in the embodiments of this application can obtain the target image quality index input by the user, and the prediction model can back-deduce the most matching detection process parameters based on the target image quality index. It does not need to rely on expert experience, improves optimization efficiency, and the target image quality index can include multiple indicators, which can realize multi-objective optimization to improve image quality.

[0145] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0146] The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0147] When the processor 602 executes the program, it implements the training method for the digital X-ray imaging detection prediction model or the optimization method for digital X-ray imaging detection process parameters provided in the above embodiments.

[0148] Furthermore, electronic devices also include:

[0149] Communication interface 603 is used for communication between memory 601 and processor 602.

[0150] The memory 601 is used to store computer programs that can run on the processor 602.

[0151] The memory 601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0152] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0153] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0154] The processor 602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0155] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0156] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0157] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0158] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0159] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A training method for a digital ray imaging detection prediction model, characterized in that, Includes the following steps: A prediction model for digital X-ray imaging detection is constructed, comprising an input layer, an output layer, a hidden layer, and an activation function. The hidden layer determines a nonlinear region based on the magnitude of the output value change of the prediction model, and sets center points according to the nonlinear region. The activation function uses a mixture of Gaussian kernel and inverse radial kernel function to fit the nonlinear region. The process of setting the center points includes: selecting an initial center point based on a sample density weighting strategy; identifying target samples whose output value change magnitude is greater than a preset magnitude, and increasing the weight of the target samples to increase the number of center points of the hidden layer in the nonlinear region. Obtain an orthogonal experimental dataset for digital ray imaging, generate a training dataset based on the orthogonal experimental dataset, train the prediction model using the training dataset, and optimize the nodes of the hidden layer based on the mean square error and prediction accuracy of the samples in the training dataset during the training process. A test set is constructed using randomly combined parameters. The prediction model is then tested using the test set. The network parameters of the prediction model are updated based on the test results until the prediction performance of the prediction model meets the target requirements.

2. The training method for the digital ray imaging detection prediction model according to claim 1, characterized in that, The weights of the target samples are: ； in, For the first The weights of each target sample, For the first The magnitude of change in the output value of each target sample For the first The magnitude of change in the output value of each sample The total number of samples, For sample index variables; The loss function of the prediction model is: ； in, The total number of samples, For the first The weights of each sample, For the first The predicted value for each sample, For the first The true value of each sample This represents the total loss of the prediction model.

3. The training method for the digital ray imaging detection prediction model according to claim 1, characterized in that, After setting the center point according to the nonlinear region, the method further includes: Identify the distribution density of samples near each center point; Adjust the width parameter of the corresponding center point based on the distribution density.

4. The training method for the digital ray imaging detection prediction model according to claim 1, characterized in that, Training the prediction model using the training dataset includes: The prediction model is trained using samples from the training dataset, and the mean squared error of the prediction model is obtained. The local gradient of the prediction model is calculated based on the mean square error; The center point and center point width parameters of the hidden layer are adjusted using the local gradient.

5. The training method for the digital ray imaging detection prediction model according to claim 1, characterized in that, The acquisition of the orthogonal experimental dataset for digital ray imaging includes: Acquire multiple input parameters for the digital X-ray imaging, wherein the input parameters are detection process parameters; A target orthogonal array is determined based on the multiple input parameters and multiple levels, and multiple sets of input parameters are combined based on the target orthogonal array. Digital ray imaging is performed on the multiple sets of input parameters to obtain corresponding output parameters, wherein the output parameters are image quality evaluation indicators; The orthogonal experimental dataset is constructed based on the multiple sets of input parameters and corresponding output parameters.

6. A method for optimizing process parameters in digital X-ray imaging detection, characterized in that, Optimizing the prediction model trained using the training method for the digital ray imaging detection prediction model as described in any one of claims 1-5 includes the following steps: Obtain the target image quality metrics input by the user; Based on the target image quality index and the prediction model, optimize the digital X-ray imaging detection process parameters that conform to the target image quality index.

7. A training device for a digital ray imaging detection prediction model, characterized in that, include: A construction module is used to construct a prediction model for digital X-ray imaging detection. The prediction model includes an input layer, an output layer, a hidden layer, and an activation function. The hidden layer determines a nonlinear region based on the magnitude of the output value change of the prediction model and sets center points according to the nonlinear region. The activation function uses a mixture of Gaussian kernels and inverse radial kernels to fit the nonlinear region. The process of setting the center points includes: selecting an initial center point based on a sample density weighting strategy; identifying target samples whose output value change magnitude is greater than a preset magnitude, and increasing the weight of the target samples to increase the number of center points of the hidden layer in the nonlinear region. The training module is used to acquire an orthogonal experimental dataset for digital ray imaging, generate a training dataset based on the orthogonal experimental dataset, train the prediction model using the training dataset, and optimize the nodes of the hidden layer based on the mean square error and prediction accuracy of the samples in the training dataset during the training process. An update module is used to construct a test set using randomly combined parameters, test the prediction model using the test set, and update the network parameters of the prediction model based on the test results until the prediction performance of the prediction model reaches the target requirements.

8. A device for optimizing process parameters of digital X-ray imaging inspection, characterized in that, Optimizing the prediction model trained using the training device for the digital X-ray imaging detection prediction model as described in claim 7 includes: The acquisition module is used to acquire the target image quality indicators input by the user; An optimization module is used to optimize digital X-ray imaging detection process parameters that conform to the target image quality index based on the target image quality index and the prediction model.

9. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the training method for the digital X-ray imaging detection prediction model as described in any one of claims 1-5, or the method for optimizing digital X-ray imaging detection process parameters as described in claim 6.

Images